Tired of your AI chatbot’s constantly-glazing therapy-speak? You could instead try striking up a conversation with “Talkie,” an old-timey AI model which is trained purely on books, newspapers, and other text sources from before the year 1930.
With its thirteen billion parameters, the researchers behind Talkie say it’s the largest “vintage” model they’re aware of, capable of holding down a conversation as if truly stuck in a past when movies with sound in them were still a novel phenomenon, and when news announcers rattled off the latest signs of tumult in the world in a bouncy Mid-Atlantic accent.
Intriguingly, Talkie is “basically” unaware of the fact it’s limited to pre-1930 times, according to David Duvenaud, an associate professor of computer science at the University of Toronto. The AI, he explained in a tweet, “doesn’t have a system prompt and they’re not smart enough yet (as far as we can tell) to introspect well enough to figure out their cut-off date.”
Talkie isn’t perfect. The researchers note that it exhibits signs of “temporal leakage,” in which it produces clearly anachronistic answers, such as knowing that “Franklin D. Roosevelt was president of the United States from 1933 to 1937.” This shows the difficult of keeping its data set pure.
Nonetheless, it raises fascinating questions. What is an LLM’s ability to predict the future? Can the nearly-century-old AI learn a modern programming language? Better yet, can it make scientific discoveries?
“As Demis Hassabis has asked,” the researchers wrote in a blog post, referring to the Google DeepMind CEO, “could a model trained up to 1911 independently discover General Relativity, as Einstein did in 1915?”
These questions remain hazy, but early tests showed that Talkie was able to create one-line programs, but still has “a long way to go before this capability is notable.”
It’s also not quite a soothsayer: tests performed by the researchers showed that the AI found historical events summarized in the New York Time‘s “On This Day” section to be more “surprising” after the knowledge cutoff, especially in the 1950s and 1960s.
That doesn’t mean its extrapolations aren’t amusing. In one user’s testing, Talkie predicted that another World War would break out in 1936, and that “flying machines” would be in everyday use for transport. Though it also strangely predicts that by 1999, “the sun will have ceased to shine,” perhaps reflecting contemporaneous anxieties over the dawn of a new millennium.
In another test by the researchers, Talkie called talking pictures — the slang for them being its namesake — “overrated.”
“They will never replace silent films, but may supplement them, and possibly in the future they may be shown in the same theater at the same time,” it confidently predicted in its long-winded style. “At present, however, they are interesting chiefly as a novelty.”
More on AI: Police Are Using AI Camera Networks to Stalk Women
